[Numpy-tickets] [NumPy] #355: Valgrind warning on arrayobject.c

NumPy numpy-tickets at scipy.net
Sat Nov 4 07:16:41 CST 2006


#355: Valgrind warning on arrayobject.c
--------------------+-------------------------------------------------------
 Reporter:  faltet  |        Owner:  somebody
     Type:  defect  |       Status:  new     
 Priority:  normal  |    Milestone:          
Component:  Other   |      Version:  devel   
 Severity:  minor   |   Resolution:          
 Keywords:          |  
--------------------+-------------------------------------------------------
Comment (by faltet):

 I've finally tried this out. I was getting the same problem with r3428.

 I've made some research, and it seems that this only appears when using
 arrays coming from numarray and converted to !NumPy objects using the
 array protocol. If I disable the numarray package in !PyTables tests (i.e.
 conversions through the array protocol doesn't occur), this problem
 doesn't show up.

 I don't know for sure, but maybe the numarray implementation of the array
 protocol is flawed (using Numeric for doing conversions seems to work
 well). So, by judging the valgrind message, the most probable thing that
 is happening is that descr->elsize is not being initialized under some
 situations (like converting special numarray arrays to numpy).

 Sorry for not being able to provide an example of that (the code is too
 messy, and I'm too lazy to look into it). One possibility is that this
 problem would be related with:

 http://sourceforge.net/tracker/index.php?func=detail&aid=1565683&group_id=1369&atid=450446

 because this example was extracted from a similar test.

 Anyway, provided that the problem is not in !NumPy itself (unless the
 problem could be a defect in !NumPy implementation of the array
 interface), maybe the best should be to close this ticket.

 Regards,

-- 
Ticket URL: <http://projects.scipy.org/scipy/numpy/ticket/355#comment:2>
NumPy <http://projects.scipy.org/scipy/numpy>
The fundamental package needed for scientific computing with Python.


More information about the Numpy-tickets mailing list